Principles
Designed to fail closed.
The decision engine is built around a small set of non-negotiable rules. None of them make the system more permissive.
The five rules
What never changes, regardless of mode.
Deterministic evaluation
Identical inputs produce identical decision semantics. There is no model sampling in the decision path.
Evidence before approval
Missing required evidence cannot silently become an allow. It becomes a named, reportable gap.
Unknown states are not assumptions
Unsupported or unrecognized actions default to denial or escalation, never to a best guess.
Authority cannot expand itself
Mandates and approval boundaries are explicit and versioned. Nothing in the decision path can widen them.
Human review cannot self-approve
The actor who proposes an action is not permitted to be the actor who approves it.
Where this shows up
Enforced in the architecture, not just stated here.
Each rule above maps to a specific layer of the decision pipeline — see the full architecture. The control layer is where mandates, policies and authority boundaries are checked; the decision layer is where an unrecognized action class resolves to denial rather than a default allow.
See these rules applied to a real decision.
Commercial analysis walks through how a supplier decision is evaluated against these same constraints.